BRN Discussion Ongoing

[CHIPS]

Something fascinating is happening inside the world’s data centers.
For more than a decade, GPUs have powered the AI revolution.

But as models grow to trillions of parameters and power costs surge, we’re hitting physical limits — and the race is on to reimagine what computing hardware should look like.

What’s emerging is nothing short of a quiet revolution — new architectures that might finally challenge the GPU’s monopoly on AI compute.

⸻

Cerebras Systems

Building wafer-scale engines — chips the size of dinner plates that can train massive models without the pain of GPU clusters. One chip, one model, no fragmentation.

Groq

Taking a radical approach with a deterministic Language Processing Unit (LPU) designed for ultra-low-latency inference. Already powering sovereign AI projects in the Middle East.

Graphcore & SambaNova

Re-architecting dataflow itself — bringing compute and memory closer together for huge efficiency gains in model training.

Lightmatter

Computing with light instead of electricity.
Its photonic chips use optical interconnects to link processors at fiber-optic speeds — a potential game-changer for hyperscale AI clusters.

BrainChip

Taking inspiration from biology, with neuromorphic chips that mimic spiking neurons to run AI at milliwatt power levels — perfect for edge AI.

Etched

Going all-in on specialization — a transformer-only ASIC that, if its claims hold, could replace hundreds of GPUs for a fraction of the cost.

⸻

Each of these players is betting on a different principle —
wafer-scale integration, deterministic execution, optical communication, or brain-like computation —
but they share one goal: break free from the GPU bottleneck.

It’s becoming clear that the future of AI compute will be heterogeneous.
GPUs will stay the workhorses, but they’ll soon be joined by optical, neuromorphic, and specialized accelerators — turning tomorrow’s data centers into orchestras of silicon.

New AI computing architectures challenge GPU dominance | Apurv Nigam posted on the topic | LinkedIn

Something fascinating is happening inside the world’s data centers. For more than a decade, GPUs have powered the AI revolution. But as models grow to trillions of parameters and power costs surge, we’re hitting physical limits — and the race is on to reimagine what computing hardware should...

www.linkedin.com

 

[Pom down under]

The Silicambrian Explosion

When BrainChip (BRN) suddenly hits mass adoption, it will spark a "Silicambrian Explosion"—a silicon/IP-fueled burst of neuromorphic innovation, mirroring the ancient Cambrian Explosion's rapid diversification of life, but this time evolving edge AI from niche experiments to everywhere-embedded intelligence. Just as the original event birthed complex multicellular life (including early brains), BrainChip's Akida chips and IP will explode into trillions of low-power, brain-like processors across IoT, autos, and beyond, turning sci-fi into silicon reality.

Buckle up: Darwin would approve of this evolutionary upgrade.

 

[Frangipani]

Thank you @Frangipani , but I do not understand the results.
Can you explain further, or is there a page missing? The last page, "Experimental Results", does not state which chip it is.

Guten Abend, CHIPS,

the abbreviations DPU, TPU and NPU refer to three different hardware platforms, cf. the preceding presentation slide titled “AI Model Overview” or the following excerpt from the May 2025 post of mine I had tagged above:

https://thestockexchange.com.au/threads/brn-discussion-ongoing.1/post-462394

“The revised GIASaaS (Global Instant Satellite as a Service, formerly Greek Infrastructure for Satellite as a Service) concept website by OHB Hellas is finally back online - and Akida now no longer shows up as UAV1 (see my post above) but as one of three SATs! 🛰

AKIDA, the envisioned SAT with the AKD1000 PCIe Card onboard, is slated to handle both Object Detection as well as Satellite Detection, while the planned KRIA and CORAL satellites (equipped with a Xilinx KRIA KV260 Vision AI Starter Kit resp. a Google Coral TPU Dev Board) are tasked to handle Vessel Detection, Cloud Detection and Fire Detection (for some reason OHB Hellas removed Flood Detection from the list of applications).

Please note that this is currently a concept website only.”


Flood detection was originally also slated to be tasked by both KRIA (DPU) and CORAL (TPU) (see here as well as under Experimental Results), but is no longer listed as a choosable application on the updated GIASaaS website https://giasaas.eu/.


Generally speaking, it would of course be helpful to also have the video recordings to go along with these conference presentations…